How do I set AI up to help me run a garden project instead of starting from scratch every time?
Before a single measurement, a single quote or a single spade, we lost weeks to a problem nobody warns you about: the AI has no idea what project it is on, and by default it never will.
What it was costing us
Around twenty minutes of every single session spent re-explaining the same garden, and decisions made on Monday quietly contradicted on Thursday because nothing was written down anywhere the AI could reach.
What we were trying to do
Run a garden project with AI as the working partner rather than as a picture generator. Brief, survey, design, planning, sequencing, budget, build — the whole thing, over months, with two of us who have never run a construction project.
What we assumed was that the hard part would be the garden. It was not. The hard part arrived on day one and had nothing to do with plants.
What we tried
We opened a chat and described the garden. Good answers. The next day we opened another chat and described the garden again, slightly differently, and got slightly different answers.
So we started pasting a summary at the top of each session. That worked for about a fortnight, until the summary was long enough to be annoying and we began trimming it. Then we tried keeping everything in one enormous conversation, which worked until it became impossible to find anything in it and the early context stopped carrying weight. We turned on the memory feature and found it had cheerfully retained that we like a warm garden with a television in it, and nothing whatsoever about the dimensions.
Then we made it worse. We use two AI tools on this project — Claude, in Cowork, and Codex inside ChatGPT — because they are good at different things, and neither could see anything the other had been told. Every fact we established in one was invisible to the other. We were not running one project with a bad memory; we were running two projects that occasionally agreed.
Meanwhile the actual project state — what we had decided, what we had rejected and why, which numbers were measured and which were guessed — was living in our heads and in a scatter of screenshots.
What went wrong
Nothing broke. That is the whole difficulty.
We settled the same questions repeatedly without noticing, because each answer was individually reasonable. A decision made on Monday was quietly contradicted on Thursday, and neither of us caught it, because we were not comparing Thursday to Monday — we were just reading Thursday. About twenty minutes of every session went on re-establishing context we had already established, and the tax was invisible because it felt like normal work.
The version that finally made it obvious: we could not answer, with confidence, which numbers on our own plan had been measured with a tape and which the AI had inferred. Nobody had written it down, so nobody knew. Weeks of design were sitting on top of a question we could not answer.
Why it happens
- 1Why does the AI keep forgetting my garden?Because a chat is a conversation, not a project. When it ends, everything you established in it is gone unless you put it somewhere else yourself.
- 2Why doesn't it just remember, when it clearly remembers some things?Consumer memory features store scattered facts about you, not a structured model of a project. They are built to make the next chat feel familiar, not to hold a brief, a set of measurements and a decision history that has to stay internally consistent for months.
- 3Why does that matter more than it sounds?Because you do not notice it failing. You notice a slightly different answer, and you assume you asked the question differently. The project drifts rather than breaks, and drift is much harder to spot than an error.
- 4Why can't I just paste my notes in each time?You can, and it works, until the notes get long enough that you start trimming them to save time. What you trim is whatever has not caused a problem recently, which is exactly the constraint that is about to.
- 5So why has nobody solved this properly?Because the tools are built around the conversation as the unit of work. A project spanning months, with a fixed brief, immovable facts and a decision history, is a different shape entirely, and nothing in the consumer products holds that shape for you.
What actually works
Stop treating the chat as where the project lives. Give the project somewhere of its own, and open each session by pointing the AI at it.
And put it somewhere every tool can read. This is the part we got wrong for longest. We use Claude in Cowork and Codex in ChatGPT on the same project, and the moment the project state lives inside either one's memory feature, the other is working blind. Plain documents in a folder solve it completely: both read the same files, both write to the same decisions log, and there is exactly one version of the truth rather than one per vendor. It also means that when we change tools next year — and we will — the project does not have to be re-explained to whatever replaces them.
In practice that is four things, and none of them need code. A brief: what this is for, what it must achieve, what is fixed and cannot be redesigned. A decisions log: dated, with the reason and the options rejected, so a settled question stays settled. A source register: every photo, quote and plan you have relied on, numbered, so any claim can be traced back. A project memory the AI reads at the start of a session, rather than a summary you retype.
Then one habit that costs nothing: when the AI tells you something that changes the project, that goes in the log before you act on it. Not afterwards, when you have forgotten the reasoning.
The test for whether you have done it right is uncomfortable and quick: open a different AI tool, point it at the folder, and ask it something about the project. If it answers the same way, the project lives in the documents. If it flounders, the project was living in a conversation and you had not noticed.
The effect is not that the answers get cleverer. It is that they stop contradicting each other, and that you can tell the difference between something you established and something the AI supplied. That distinction turns out to be the whole game — it is what separates a plan you can build from a plan that merely reads well.
We document this project the same way we run it, which is not a coincidence. Everything on this site comes out of that same record: the problems, the decisions, the sources. If the method did not survive being written down in public, it would not be much of a method.
- This is an organising method, not a project management qualification. It will not tell you whether your sequence of works is right.
- It assumes you are willing to write things down once. If nothing ever gets recorded, nothing can be remembered.
- It does not make the AI's answers correct. It makes them consistent, which is a different and lesser thing.
- Everything here is deliberately consumer-grade: documents and folders, no code, no developer tooling.
Words used above
- Brief
- One document holding what the project is for, what it must achieve and what is fixed. The thing every later answer has to agree with.
- Decisions log
- A dated list of what was decided and why, including the options rejected, so a settled question stays settled.
- Source register
- A numbered list of every photograph, quote, plan or page you have relied on, so a claim can be traced back to where it came from.
- Project memory
- Somewhere outside the chat window that holds all of the above, and which the AI can read at the start of a session instead of being told again.
The explanation is free. The instrument is paid.
You now know why it fails. Here is what fixes it.
- BriefingThe One-Page Garden Brief — The fifteen things to write down before you ask AI anything, in the order that stops the first answer being useless. Fill it in once and paste it at the top of every session.
- CourseSet Up a Project Brain AI Can Actually Use — Ninety minutes, ending with a working brief, decisions log and source register for your own project, plus the session-opening routine that keeps the AI on the rails. No code, no terminal.
- BlueprintBuild a Project Memory for Domestic Builds — The full build spec for the tool we wished existed: structured project memory an AI can read directly, with a brief, immovable facts, decisions and sources. Screens, data model and acceptance tests.
Free trial, then paid. Everything above is inside The AI Garden Project.
Published 25 July 2026